Canada

Large language models are a fascinating and powerful technology, but are not faultless. Iteratively selecting the most probable next word can result in a sentence that sounds coherent, but contains content that is fabricated, taken out of context, or in self-conflict. This can create particular danger when end users rely on LLMs for information on critical issues. Many countries lack equal or affordable access to healthcare. This can lead to some viewing popular LLM chatbots such as ChatGPT as being an online doctor; available to answer their healthcare needs. People with chronic diseases are particularly vulnerable to this, as the chronicity of their disease presents a consistent healthcare cost. Our research seeks to take the first step of many in creating a framework within which LLMs can self-check for hallucinations, resulting in more reliable information. We worked in reference to an existing paper to create a dataset containing pairs of questions and hallucinated answers, specifically in reference to chronic disease. We also conduct an experiment on our dataset in which we determine whether LLMs can accurately detect their own hallucinations. We find that, while some LLMs can identify a hallucinated answer with incredible accuracy, it is highly inconsistent from model to model. We endorse the future research of hallucination detection technologies in LLMs in order to mitigate risk of harm caused by the dissemination of inaccurate information by LLMs.

Indicium 2025

A Reflection on Hallucinations in LLM Responses to Chronic Disease Queries

large language model

hallucination

chronic disease

**Welcome to EMNLP 2023!**

On behalf of the EMNLP 2023 Organizing Committee, I extend a warm and heartfelt welcome to all of you to the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). It is with immense pleasure and excitement that we gather here in Singapore, a vibrant hub of innovation and technological advancement.

The conference program is packed with insightful presentations, thought-provoking workshops, and engaging networking opportunities. In addition to the technical sessions, we have also planned several social events that will provide you with the opportunity to connect with your colleagues.

The conference is held in person at the Resorts World Convention Centre in Singapore, and available on-line with the help of Underline.

For those in Singapore, I hope that you will find time to explore the exciting Sentosa Island, Gardens by the Bay, Singapore Botanic Gardens and the many other attractions unique to Singapore.

Of course, we are grateful to our sponsors and partners for their generous support of EMNLP 2023, whose contributions make it possible for us to host this world-class event.

Yuji Matsumoto (RIKEN AIP) 
EMNLP 2023 General Chair

To access the **EMNLP 2023** event page on Underline, you need to register for the Conference. 
Please follow **[this link](https://2023.emnlp.org/registration/)** for more details.

EMNLP 2023

EMNLP 2023 took place in Singapore from Dec 6th to Dec 10th, 2023.

technical paper

This annual scholarly program is designed to provide undergraduate students with a structured introduction to laboratory research within their home institution. Participants are matched with principal investigators and mentored by graduate students, enabling them to engage directly in ongoing research projects and develop foundational skills in scientific inquiry.

This annual scholarly program is designed to provide undergraduate students with a structured introduction to laboratory research within their home institution.
This year’s cohort focused on projects in the field of longevity science, exploring diverse topics such as bioinformatics, molecular biology, and cognitive science.

Fragile X syndrome (FXS) is a neurodevelopmental disorder caused by the expansion of cysteine-guanine-guanine (CGG) repeats in the fragile X messenger ribonucleoprotein 1 (FMR1) gene. The increase in repeats silences the transcription of fragile X messenger ribonucleoprotein (FMRP), resulting in intellectual and developmental disabilities. Despite the development of symptom management therapies, there is no cure for FXS. In the past decades, several studies have demonstrated the importance of age-related alterations in the processing of amyloid precursor protein (APP) in FXS pathology. Hence, this systematic review was conducted to summarize and evaluate FXS treatments targeting components of the APP pathway. Using PubMed, a literature search was performed and uncovered 88 results. 9 articles related to the APP pathway were selected, and they were further categorized by their targets: glutamate, metabotropic glutamate receptor 5 (mGluR5), brain cytoplasmic RNA (BC1), and A disintegrin and metalloproteinase 10 (ADAM10). The majority of these results focused on treating the symptoms in young patients, corresponding with peak symptomatology. However, FXS individuals also display age-dependent premature cognitive decline. Since APP processing is altered with age, we suggest these treatments for FXS in younger individuals may be translatable to the aging FXS population. Further research is required to understand the mechanism of age-dependent alternation in APP processing in FXS patients and to investigate the efficiency of treatments with increased sample size and different tissues.

Aging with Fragile X Syndrome: Prospective Treatments Targeting the Amyloid Processing Pathway

People from around the world experience mobility impairments, leading to significant impacts on independence and quality of life. While conventional mobility assistive devices such as wheelchairs, walkers, and prosthetics offer essential support, they often lack adaptability and user responsiveness. Artificial intelligence redefines mobility devices by advancing autonomy, personalization, and user experience. This scoping review maps the current landscape of AI-powered mobility aids for adults with physical impairments. Using the Arksey and O’Malley framework and PRISMA-ScR guidelines, we systematically reviewed 39 peer-reviewed studies, which were identified through specific search strategies developed for each database. Screening and data extraction were conducted in pairs using a predefined criterion. The included studies spanned 17 countries and focused primarily on smart wheelchairs (n=22), followed by walkers (n=8), prosthetics (n=3), rollators (n=2), and individual studies on canes, wheelchairs/scooters, lower-limb braces, and body-weight support systems (n=1 each). AI models were then categorized by learning type. AI-powered mobility devices offer individual benefits like enhanced navigation, motivation, and emotional well-being, but users face challenges such as device complexity, usability issues, and fatigue. At the administrative level, these technologies reduce caregiver workload and improve efficiency, yet are hindered by insufficient training and limited user input, while policy-level barriers include inadequate funding and lack of regulatory support. These findings affirm the potential of AI to enhance mobility and autonomy and promote the urgent need for inclusive, real-world research, standardized evaluation, and equitable access across diverse populations.

AI-powered Mobility Devices for Adults With Impairments: A Scoping Review

As populations age globally, older adults are increasingly affected by chronic physical and mental health challenges, including frailty, cognitive impairment, depression, and social isolation. Artificial intelligence (AI) tools such as chatbots, wearable monitors, predictive models, and socially assistive robots have been proposed as scalable solutions to support aging populations. This literature review systematically examined 28 peer-reviewed studies to evaluate the efficacy and usability of current AI interventions designed to enhance health and well-being in adults aged 65 and older. Methods included supervised machine learning for diagnosis and prediction, large language models for health communication, and co-designed systems for home-based care and emotional support. Results suggest that AI applications can improve physical activity, emotional regulation, and early detection of cognitive decline and depression, with user feedback indicating positive perceptions of usability and engagement. However, most studies had small, homogeneous samples, lacked long-term follow-up, and were conducted in simulated or non-clinical settings. Ethical and implementation challenges were common, including data privacy concerns, unclear algorithmic decision-making, limited digital literacy, and underrepresentation of diverse older adult populations. Despite these limitations, the evidence indicates that AI holds promise in augmenting and streamlining older adult care when tools are transparent, inclusive, and tailored to user needs. Broader adoption will require rigorous evaluation in real-world environments, equitable design practices, and sustained attention to ethical risks.

Artificial Intelligence for Care for Older Adults with Chronic Health Conditions

Loneliness is increasingly recognized as a non-motor symptom in Parkinson’s disease (PD), impacting mood, cognitive decline, and even mortality. Individuals with PD are vulnerable due to physical limitations, communication challenges, and stigma. This review synthesizes peer-reviewed literature on the intersection of loneliness, social isolation, and PD. Using Covidence, we screened 232 abstracts and subsequently reviewed 102 full-text articles. Twenty contributed contextual or descriptive insights, while only four studies provided structured, analyzable data using validated measurement tools or reporting statistical associations with clinical outcomes. Most studies were observational, published between 2020 and 2024, in English. Where reported, loneliness prevalence in PD populations ranged from 24% to over 30%, and was linked to poor quality of life, depression, and a greater disease burden. One prospective study identified a significant association between loneliness and incident PD (HR=1.25, fully adjusted), while others connected loneliness to reduced autonomy and poorer functional outcomes. However, intervention studies aimed at alleviating loneliness were notably rare and typically descriptive, with no high-quality trials identified. A concern was the limited use of formal risk bias tools, with only 17% of studies employing one. Furthermore, conflict of interest disclosures were present in a mere 13% of the articles included. While 60% of studies did include reviewer's comments on bias, these lacked standardization. This review underscores a growing but methodologically inconsistent body of literature. Improved standardization is crucial to develop evidence-based targeted interventions to address loneliness in individuals with PD.

Bridges in Parkinson’s Disease: Connecting Lives, Ending Loneliness

As the global economy expands, its environmental consequences intensify. Atmospheric CO2 now exceeds 420 ppm and continues rising. Greenhouse gases use CO2 enrichment to boost crop yields, but most rely on fossil fuel-derived CO2. This study models a Direct Air Capture (DAC) system using Zeolite 13X to recycle CO2 for greenhouse enrichment. Regeneration via solar, electricity, and natural gas was compared to determine which energy source offers the most sustainable outcome. A custom gas chamber was packed with Zeolite 13X and outfitted with four thermocouples. Following calibration, the device was connected to a carbon dioxide pump to measure and analyze the total adsorption level. Zeolite 13X effectively captured and released CO2 using all regeneration methods. The highest CO2 adsorption was observed using electric heat wrap (0.275 mmol CO2/g) and solar treatment (0.201 mmol CO2/g). Adsorption was greater in heat-assisted trials with Zeolite 13X compared to controls. Energy inputs remained consistent across sources: grid electricity (0.2653 MJ/kg CO2), solar (0.2642 MJ/kg CO2), and natural gas (0.2681 MJ/kg CO2). Grid electricity produced the lowest emissions per kilogram of CO2 captured (0.003 kg CO2 emitted/kg CO2), followed by solar (0.008 kg/kg), and natural gas (0.013 kg/kg). The experimental findings validate the energy-efficient nature of Zeolite 13X carbon capture technology which shows promise for agricultural and clean energy sector applications.

CO₂ Adsorption and Regeneration of Zeolite 13X: Energy Resource Comparison

Heart disease remains the leading global cause of death, highlighting the urgent need for accurate and scalable early detection methods. This study investigates the effectiveness of supervised machine learning techniques in predicting heart disease from routine clinical data. The objective is to determine which classification model provides the most reliable predictions in a clinical setting. We combined four public heart disease datasets from the UCI Machine Learning Repository and selected 14 relevant clinical features. Missing values were filled in using standard statistical methods (mean, median, or mode as appropriate). We then trained and evaluated four supervised machine learning models: logistic regression, support vector machine, extreme gradient boosting, and a multi-layer neural network. Each model was tuned for optimal performance and assessed using accuracy, precision, recall, and the F1-score. All models achieved higher than 80% accuracy. The support vector machine was the most accurate with 83.2% on the test data and also attained the highest F1-score (73%) for identifying heart disease cases. The neural network and gradient boosting models each reached about 82% accuracy, with the neural network identifying slightly more actual positive cases (higher recall). The logistic regression model trailed slightly at 80.7% accuracy but still showed competitive precision and recall. Machine learning models can effectively predict heart disease, providing a valuable tool for early detection. Among the models tested, the support vector machine performed best. These results indicate that support vector machines may offer stronger predictive performance for clinical applications of heart disease detection.

Evaluating Supervised Machine Learning Models for Heart Disease Prediction

Optical Coherence Tomography (OCT) is a widely adopted imaging technique in ophthalmology, yet its clinical interpretation remains prone to inconsistency and subjectivity. This study presents a retrieval-augmented generation (RAG) framework aimed at improving both diagnostic accuracy and interpretability in OCT-based analysis. The pipeline integrates a dual-headed convolutional neural network with a case-based retrieval system and a language model capable of generating structured clinical summaries. A ResNet-50 model was trained to classify diabetic retinopathy (DR), diabetic macular edema (DME), and key retinal biomarkers using a public OCT dataset. Feature embeddings generated by the model were indexed into a similarity-search database to enable retrieval of clinically relevant cases. Retrieved scans, along with classification outputs, were used to prompt a large language model to generate concise diagnostic reports that reflect both the model’s predictions and visually aligned precedents. The system achieved 89% accuracy in distinguishing DR from DME in a hold-out test set. Interpretability was also demonstrated through consistent retrieval of semantically similar cases. Generated reports were found to be readable and aligned with clinical reasoning. By combining classification performance with transparent, case-grounded explanations, the RAG approach enhances clinical confidence and supports decision-making. This work highlights the potential of retrieval-augmented pipelines to reduce variability in OCT interpretation and to serve as a scalable, explainable foundation for future AI-driven diagnostic support in ophthalmology.

Improving Optical Coherence Tomography Interpretation With Retrieval-Guided Diagnosis and Text Generation

As global energy demand rises and climate concerns intensify, solar energy systems or photovoltaic systems are becoming increasingly vital sources of renewable energy. However, their potential remains limited by the inability to operate naturally at the Maximum Power Point during hours of fluctuating sunlight. This study addresses this challenge by improving photovoltaic system efficiency through advanced Maximum Power Point Tracking techniques. We hypothesized that real-time duty cycle control would allow actual power to consistently rise towards Pmax, regardless of fluctuations. Using simulation software PowerSIM (PSIM), we analyzed the impact of varying duty cycles on power output across several trial runs. We conducted simulations on four PV panels, each with a distinct maximum power value, by modifying the duty cycle under five controlled power fluctuations designed to mimic variable sunlight. A code was developed and implemented to dynamically adjust the duty cycle based in response to changes in output power. The results showed that the PV panels with the highest (395W) and lowest (225W) Pmax values responded effectively, with actual power rising towards Pmax after each fluctuation. To validate our findings, we implemented the algorithm on a digital signal processing board using PSIM’s Processor-in-the-Loop module. The test completed in 4 minutes and 27 seconds, with the system's duty cycle varying with changes in actual power within a 0.03 range. The broader significance of this project lies in its contributions to developing responsive and cost effective PV systems that align with the goals of economic resilience and sustainable chronic disease management.

Improving Photovoltaic Systems for Renewable Energy Efficiency and Affordably Powering Aging Communities

This paper illustrates the process of transforming an initial mathematical permutation of numbers into a final permutation through a series of conditional swaps. The goal is to explore various approaches for modeling and analyzing all valid swap sequences that can achieve such a transformation. To address this and to better understand these transformations, different mathematical and combinatorial representations need to be studied, in order to gain useful insight into the underlying patterns of these transformations. This paper outlines a framework for understanding permutation transformations, solving problems modeled by similar rules, and provides grounds for further research in combinatorics and optimization.

Modelling Mathematical Permutation Transformations: Using Canonical Chains and Combinatorial Objects

Osteoarthritis (OA) is a prevalent and disabling condition affecting over one third of Canadians aged 65 and older, with the hand and wrist being among the most impacted joints. This chronic, degenerative joint disease severely limits physical independence by causing synovial inflammation and cartilage degradation, leading to joint stiffness. As a result, seniors diagnosed with OA often experience persistent joint pain, progressive joint immobility, and diminished quality of life. Current treatments such as pharmacotherapy and physiotherapy often provide limited or inconsistent relief due to side effects, accessibility, or patient adherence. In response, this project introduces a novel, multimodal therapeutic glove designed to deliver lasting non-pharmacological relief for individuals with hand and wrist OA. The glove integrates thermal stimulation and localized vibration to reduce pain and joint immobility, improving hand function. The design is grounded in an evidence-based understanding of osteoarthritis pain responses, with therapeutic principles derived from case studies and meta-analyses of both thermotherapy and massage therapy. The glove features vibrating mechanisms that simulate massage movements, providing relief and promoting circulation in affected areas. Furthermore, the glove incorporates heat therapy, which helps reduce stiffness and alleviates pain by increasing blood flow. To evaluate the glove’s effectiveness and usability, a pilot test was conducted with nine participants, where two-thirds of the participants reported hand or wrist pain. Guided sessions were used to measure pain and assess mobility using a numerical rating scale (0–5). The average reported pain relief score was 3.89, while the ease-of-movement score averaged 4.56, suggesting moderate to significant improvement.

Downloads

Next from Indicium 2025

Aging with Fragile X Syndrome: Prospective Treatments Targeting the Amyloid Processing Pathway

Similar lecture

HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models | VIDEO